0-dimensional ndarrays should support len

[gerritholl]

When asking for len(X) on a 0-d ndarray, a TypeError is raised because the object is reportedly unsized. This despite X.size being equal to 1. If X were unsized, it wouldn't have a '.size' attribute?

In [88]: X = array(0.0, dtype=float32)

In [89]: X.shape
Out[89]: ()

In [90]: X.size
Out[90]: 1

In [91]: len(X)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-91-a6c197cecae8> in <module>()
----> 1 len(X)

Interestingly, a void does accept len:

In [80]: X = zeros(shape=(), dtype=dtype([('x', '<f4')]))

In [81]: X.size
Out[81]: 1

In [82]: len(X)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-82-a6c197cecae8> in <module>()
----> 1 len(X)

TypeError: len() of unsized object

In [83]: len(X[()])
Out[83]: 1

[rkern]

It does not report a __len__ because it cannot be indexed by integers, sliced, or iterated over, at least one of which should be true for things with a __len__. .size is a numpy thing talking about how much memory it takes up (in units of .itemsize bytes), more or less.

[gerritholl]

Or maybe not. I just thought the second example was odd. If it has a size, doesn't that sort of mean it has a length?

[gerritholl]

It makes sense, on second thought. But then why does void have a __len__?

[njsmith]

Another way to think about it: a.size is prod(a.shape), and len(a) is a.shape[0]. So for a 0d array, the former is 1 and the latter is an error.

[rkern]

This void is a record with fields. You can access those fields by indexing in addition to using the name.

[seberg]

I guess this can be closed, or is there still something to discuss here @gerritholl?